The Working Memory Model

In brief

In 1974, Alan Baddeley (born 1934) and Graham Hitch published a chapter titled "Working Memory" that dismantled a picture then dominant in cognitive psychology: that short-term memory was a single, unitary store through which nearly all information passed en route to long-term storage. They proposed instead a system of parts: a central executive directing attention, and two specialized helpers under it, a phonological loop for sound and language and a visuospatial sketchpad for images and layout. In 2000 Baddeley added a fourth component, the episodic buffer, to explain how these streams bind into one coherent experience. The new name mattered as much as the new structure: no longer passive storage, but a workspace for holding and manipulating information in the service of a task.

The full treatment

The problem it answers

The standard account, laid out by Richard Atkinson and Richard Shiffrin in 1968, treated memory as a sensory register feeding a short-term store feeding a long-term store, with rehearsal moving information from the second into the third. A key case strained that picture: if the short-term store were the sole gateway to long-term memory, patients with badly damaged short-term memory should struggle to learn anything new. Tim Shallice and Elizabeth Warrington's 1970 study of a patient known as KF found the opposite. His auditory memory span fell to about two items, yet he learned new information normally, and his memory for visual material was intact, a split the unitary model could not produce.

How it works

The central executive is an attention-controlling system with no storage of its own: it directs attention, switches between tasks, and allocates work to two subordinate, modality-specific systems. The phonological loop handles sound and language: a passive phonological store holds acoustic traces for a couple of seconds before they fade, and an articulatory rehearsal process refreshes them through silent repetition and converts written material into a spoken code so it too can be held. The visuospatial sketchpad does the equivalent job for images and locations, holding visual form and layout long enough to compare shapes or plan a route; Robert Logie later split it into a visual cache for form and colour and an inner scribe for spatial movement.

What it claims

Short-term retention is not one system but several, each with its own limited capacity, able to run largely in parallel because each draws on a separate resource. Two verbal tasks compete for the same subsystem and interfere badly; a verbal task paired with a spatial one barely interferes at all. The true bottleneck is the executive itself, a general-purpose resource both subsystems draw on for control, which is why any sufficiently demanding pair of tasks still interferes, whatever their content.

The key study or demonstration

Baddeley and Hitch's original evidence came from dual-task experiments: subjects held a memory load of zero to eight digits while judging whether short sentences correctly described the order of two letters. A single-resource model predicts digit loading should badly impair the reasoning; instead, reasoning slowed only modestly and accuracy barely dropped even under the heaviest loads, exactly what separate resources predict. A second line of evidence mapped the phonological loop itself. Baddeley, Neil Thomson, and Mary Buchanan showed in 1975 that memory span for word lists tracks how many words a person can say aloud in about two seconds, the word-length effect, and that repeating an irrelevant sound during the task abolishes that effect for visual lists, because it blocks the rehearsal that would otherwise recode them into speech.

Related distinctions

"Working memory" is often used as a loose synonym for "short-term memory," but Baddeley's term is narrower: a system for holding information while it is used, not merely a duration of storage, and distinct from measures of working memory capacity such as Meredith Daneman and Patricia Carpenter's 1980 reading span task, which quantifies how much a person can hold and process at once without committing to any particular internal structure.

Lineage

The model descends from the multi-store, "modal" model of memory that Atkinson and Shiffrin set out in 1968, itself building on George Miller's 1956 observation that immediate memory span is limited to roughly seven items. Baddeley and Hitch kept the idea of a limited-capacity short-term system but broke it into parts, drawing on Donald Broadbent's earlier work treating attention as filtered and capacity-limited. Baddeley's 1986 book, Working Memory, consolidated the three-component version; his 2000 paper on the episodic buffer added the fourth, prompted by cases such as amnesic patients recalling prose far beyond what loop capacity should allow.

The strongest case for it

The model has held its place because it makes distinctions that show up cleanly in independent evidence. Double dissociations in neuropsychology support the split directly: KF and patients like him lose verbal short-term memory while sparing visuospatial memory, while other patients show the reverse. A single unitary store cannot produce a double dissociation; two separable systems can. Dual-task studies by Robert Logie, Giuliana Zucco, and Baddeley in 1990 agree: a concurrent spatial task disrupts spatial memory far more than verbal memory, and a concurrent verbal task does the reverse. The phonological loop has also proven productive outside the lab: Susan Gathercole and Baddeley showed that children's ability to repeat unfamiliar nonwords predicts how fast they learn new vocabulary, evidence Baddeley, Gathercole, and Costanza Papagno used in 1998 to argue the loop evolved as a device for learning language's sound patterns.

The strongest case against it

The most persistent objection targets the central executive. Baddeley himself acknowledged that it risked becoming a label for whatever the slave systems could not explain, a homunculus doing unspecified work rather than a mechanism, and later work, with Logie among others, tried to fractionate it into testable functions like inhibiting responses, updating information, and switching tasks.

A more direct rival comes from Nelson Cowan, whose embedded-processes model, developed from 1988 and set out in Akira Miyake and Priti Shah's 1999 volume Models of Working Memory, denies separate storage structures are needed at all. On Cowan's account, working memory is simply the activated portion of long-term memory plus a sharply limited focus of attention, holding about three to four chunks; verbal and visual material differ because they are different content, not because they sit in distinct stores, a more economical account if it fits the data.

A related challenge comes from Randall Engle, Andrew Conway, and Michael Kane, whose work on individual differences argues that what predicts fluid intelligence and comprehension is domain-general executive attention, the capacity to keep goal-relevant information active against distraction, rather than separable storage. On this view the variation between people has little to do with the loop or the sketchpad and everything to do with one attentional resource.

Where it stands now

The multi-component model remains the default account taught in introductory psychology, and its core distinctions, verbal versus visuospatial storage, storage versus executive control, have held up across five decades of evidence. But it now sits as one of several live frameworks rather than an unchallenged consensus, competing with Cowan's and Engle's accounts for the same phenomena, and its own author kept revising it, most substantially by adding the episodic buffer. A durable core surrounded by dispute over its edges is closer to how mature science usually looks than a model with no dissent at all.

Test yourself

Try holding a new phone number in mind while someone describes a route to a building you don't know. Notice how little the two tasks interfere, even though both are demanding. Now try holding that number while repeating another string of digits under your breath instead. Notice how much faster it collapses. The model predicts exactly that pattern: shared resources collapse under load, separate ones coexist.